Skip to content

Tier 1 — Proxy

What it is

A per-query estimate based on public benchmarks and model-family heuristics. Zero overhead, broadest coverage, widest intervals. The right tier for unsupported models, prototype workloads, and back-of-envelope sketches.

Inputs

  • AI Energy Score v2 (Hugging Face × Salesforce × Cohere × CMU) — model-family median energy per query, where available
  • EcoLogits priors (GenAI Impact, MPL-2.0) — emission and water priors by model family and parameter count
  • Token-ratio scaling — adjust the model-family median by the actual token count of the request
  • Regional grid intensity — JRC NEEFE 2024 annual average for the region that handled the request (ENTSO-E live not used at this tier; the proxy estimate's wider interval already covers grid variability)

Output

{
  "tier": "proxy",
  "tier_id": "01",
  "co2e": {
    "median_g": 1.62,
    "ci90": {"low": 0.85, "high": 3.10}
  },
  "pedigree": [3, 3, 2, 2, 3]
}

Pedigree expectations

Tier 1 receipts typically score [3, 3, 2, 2, 3] on the Weidema axes:

  • Reliability (3): non-verified data based on assumptions
  • Completeness (3): representative data from a smaller set of sites
  • Temporal (2): data within 6 years
  • Geographic (2): data from a similar production area
  • Technological (3): data from related processes

The pedigree score directly determines the prior dispersion that feeds Monte Carlo at tier 2. Higher scores mean wider intervals.

When tier 1 is the right answer

  • The model is not in our calibrated set (newly released model, custom fine-tune)
  • The request is in a route where tier 2 inputs are missing (rare; only on initial onboarding or after upstream provider change)
  • The customer has explicitly requested tier 1 for compatibility with their own internal accounting (we do not recommend this; tier 2 is almost always the right default)

When tier 1 is the wrong answer

  • For CSRD disclosure aggregates. Wide intervals propagate; an annual aggregate built on tier-1 receipts will fail an assurance partner's materiality check.
  • For comparative reporting against tier-2-or-higher peers. The widths are not comparable.

Where this is implemented

methodology/tier1/proxy.py

Citations

  • Hugging Face × Salesforce × Cohere × CMU. AI Energy Score v2. huggingface.co/spaces/AIEnergyScore (2025).
  • GenAI Impact. EcoLogits. github.com/genai-impact/ecologits (2024–2026).
  • Joint Research Centre. NEEFE 2024. jrc.ec.europa.eu/datasets/neefe-2024.